1 /* 2 * Copyright (C) 2013 Red Hat 3 * Author: Rob Clark <robdclark@gmail.com> 4 * 5 * This program is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 as published by 7 * the Free Software Foundation. 8 * 9 * This program is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 12 * more details. 13 * 14 * You should have received a copy of the GNU General Public License along with 15 * this program. If not, see <http://www.gnu.org/licenses/>. 16 */ 17 18 #include <linux/sync_file.h> 19 20 #include "msm_drv.h" 21 #include "msm_gpu.h" 22 #include "msm_gem.h" 23 24 /* 25 * Cmdstream submission: 26 */ 27 28 /* make sure these don't conflict w/ MSM_SUBMIT_BO_x */ 29 #define BO_VALID 0x8000 /* is current addr in cmdstream correct/valid? */ 30 #define BO_LOCKED 0x4000 31 #define BO_PINNED 0x2000 32 33 static struct msm_gem_submit *submit_create(struct drm_device *dev, 34 struct msm_gpu *gpu, struct msm_gpu_submitqueue *queue, 35 uint32_t nr_bos, uint32_t nr_cmds) 36 { 37 struct msm_gem_submit *submit; 38 uint64_t sz = sizeof(*submit) + ((u64)nr_bos * sizeof(submit->bos[0])) + 39 ((u64)nr_cmds * sizeof(submit->cmd[0])); 40 41 if (sz > SIZE_MAX) 42 return NULL; 43 44 submit = kmalloc(sz, GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY); 45 if (!submit) 46 return NULL; 47 48 submit->dev = dev; 49 submit->gpu = gpu; 50 submit->fence = NULL; 51 submit->pid = get_pid(task_pid(current)); 52 submit->cmd = (void *)&submit->bos[nr_bos]; 53 submit->queue = queue; 54 submit->ring = gpu->rb[queue->prio]; 55 56 /* initially, until copy_from_user() and bo lookup succeeds: */ 57 submit->nr_bos = 0; 58 submit->nr_cmds = 0; 59 60 INIT_LIST_HEAD(&submit->node); 61 INIT_LIST_HEAD(&submit->bo_list); 62 ww_acquire_init(&submit->ticket, &reservation_ww_class); 63 64 return submit; 65 } 66 67 void msm_gem_submit_free(struct msm_gem_submit *submit) 68 { 69 dma_fence_put(submit->fence); 70 list_del(&submit->node); 71 put_pid(submit->pid); 72 msm_submitqueue_put(submit->queue); 73 74 kfree(submit); 75 } 76 77 static inline unsigned long __must_check 78 copy_from_user_inatomic(void *to, const void __user *from, unsigned long n) 79 { 80 if (access_ok(VERIFY_READ, from, n)) 81 return __copy_from_user_inatomic(to, from, n); 82 return -EFAULT; 83 } 84 85 static int submit_lookup_objects(struct msm_gem_submit *submit, 86 struct drm_msm_gem_submit *args, struct drm_file *file) 87 { 88 unsigned i; 89 int ret = 0; 90 91 spin_lock(&file->table_lock); 92 pagefault_disable(); 93 94 for (i = 0; i < args->nr_bos; i++) { 95 struct drm_msm_gem_submit_bo submit_bo; 96 struct drm_gem_object *obj; 97 struct msm_gem_object *msm_obj; 98 void __user *userptr = 99 u64_to_user_ptr(args->bos + (i * sizeof(submit_bo))); 100 101 /* make sure we don't have garbage flags, in case we hit 102 * error path before flags is initialized: 103 */ 104 submit->bos[i].flags = 0; 105 106 if (copy_from_user_inatomic(&submit_bo, userptr, sizeof(submit_bo))) { 107 pagefault_enable(); 108 spin_unlock(&file->table_lock); 109 if (copy_from_user(&submit_bo, userptr, sizeof(submit_bo))) { 110 ret = -EFAULT; 111 goto out; 112 } 113 spin_lock(&file->table_lock); 114 pagefault_disable(); 115 } 116 117 if ((submit_bo.flags & ~MSM_SUBMIT_BO_FLAGS) || 118 !(submit_bo.flags & MSM_SUBMIT_BO_FLAGS)) { 119 DRM_ERROR("invalid flags: %x\n", submit_bo.flags); 120 ret = -EINVAL; 121 goto out_unlock; 122 } 123 124 submit->bos[i].flags = submit_bo.flags; 125 /* in validate_objects() we figure out if this is true: */ 126 submit->bos[i].iova = submit_bo.presumed; 127 128 /* normally use drm_gem_object_lookup(), but for bulk lookup 129 * all under single table_lock just hit object_idr directly: 130 */ 131 obj = idr_find(&file->object_idr, submit_bo.handle); 132 if (!obj) { 133 DRM_ERROR("invalid handle %u at index %u\n", submit_bo.handle, i); 134 ret = -EINVAL; 135 goto out_unlock; 136 } 137 138 msm_obj = to_msm_bo(obj); 139 140 if (!list_empty(&msm_obj->submit_entry)) { 141 DRM_ERROR("handle %u at index %u already on submit list\n", 142 submit_bo.handle, i); 143 ret = -EINVAL; 144 goto out_unlock; 145 } 146 147 drm_gem_object_reference(obj); 148 149 submit->bos[i].obj = msm_obj; 150 151 list_add_tail(&msm_obj->submit_entry, &submit->bo_list); 152 } 153 154 out_unlock: 155 pagefault_enable(); 156 spin_unlock(&file->table_lock); 157 158 out: 159 submit->nr_bos = i; 160 161 return ret; 162 } 163 164 static void submit_unlock_unpin_bo(struct msm_gem_submit *submit, 165 int i, bool backoff) 166 { 167 struct msm_gem_object *msm_obj = submit->bos[i].obj; 168 169 if (submit->bos[i].flags & BO_PINNED) 170 msm_gem_put_iova(&msm_obj->base, submit->gpu->aspace); 171 172 if (submit->bos[i].flags & BO_LOCKED) 173 ww_mutex_unlock(&msm_obj->resv->lock); 174 175 if (backoff && !(submit->bos[i].flags & BO_VALID)) 176 submit->bos[i].iova = 0; 177 178 submit->bos[i].flags &= ~(BO_LOCKED | BO_PINNED); 179 } 180 181 /* This is where we make sure all the bo's are reserved and pin'd: */ 182 static int submit_lock_objects(struct msm_gem_submit *submit) 183 { 184 int contended, slow_locked = -1, i, ret = 0; 185 186 retry: 187 for (i = 0; i < submit->nr_bos; i++) { 188 struct msm_gem_object *msm_obj = submit->bos[i].obj; 189 190 if (slow_locked == i) 191 slow_locked = -1; 192 193 contended = i; 194 195 if (!(submit->bos[i].flags & BO_LOCKED)) { 196 ret = ww_mutex_lock_interruptible(&msm_obj->resv->lock, 197 &submit->ticket); 198 if (ret) 199 goto fail; 200 submit->bos[i].flags |= BO_LOCKED; 201 } 202 } 203 204 ww_acquire_done(&submit->ticket); 205 206 return 0; 207 208 fail: 209 for (; i >= 0; i--) 210 submit_unlock_unpin_bo(submit, i, true); 211 212 if (slow_locked > 0) 213 submit_unlock_unpin_bo(submit, slow_locked, true); 214 215 if (ret == -EDEADLK) { 216 struct msm_gem_object *msm_obj = submit->bos[contended].obj; 217 /* we lost out in a seqno race, lock and retry.. */ 218 ret = ww_mutex_lock_slow_interruptible(&msm_obj->resv->lock, 219 &submit->ticket); 220 if (!ret) { 221 submit->bos[contended].flags |= BO_LOCKED; 222 slow_locked = contended; 223 goto retry; 224 } 225 } 226 227 return ret; 228 } 229 230 static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit) 231 { 232 int i, ret = 0; 233 234 for (i = 0; i < submit->nr_bos; i++) { 235 struct msm_gem_object *msm_obj = submit->bos[i].obj; 236 bool write = submit->bos[i].flags & MSM_SUBMIT_BO_WRITE; 237 238 if (!write) { 239 /* NOTE: _reserve_shared() must happen before 240 * _add_shared_fence(), which makes this a slightly 241 * strange place to call it. OTOH this is a 242 * convenient can-fail point to hook it in. 243 */ 244 ret = reservation_object_reserve_shared(msm_obj->resv); 245 if (ret) 246 return ret; 247 } 248 249 if (no_implicit) 250 continue; 251 252 ret = msm_gem_sync_object(&msm_obj->base, submit->ring->fctx, 253 write); 254 if (ret) 255 break; 256 } 257 258 return ret; 259 } 260 261 static int submit_pin_objects(struct msm_gem_submit *submit) 262 { 263 int i, ret = 0; 264 265 submit->valid = true; 266 267 for (i = 0; i < submit->nr_bos; i++) { 268 struct msm_gem_object *msm_obj = submit->bos[i].obj; 269 uint64_t iova; 270 271 /* if locking succeeded, pin bo: */ 272 ret = msm_gem_get_iova(&msm_obj->base, 273 submit->gpu->aspace, &iova); 274 275 if (ret) 276 break; 277 278 submit->bos[i].flags |= BO_PINNED; 279 280 if (iova == submit->bos[i].iova) { 281 submit->bos[i].flags |= BO_VALID; 282 } else { 283 submit->bos[i].iova = iova; 284 /* iova changed, so address in cmdstream is not valid: */ 285 submit->bos[i].flags &= ~BO_VALID; 286 submit->valid = false; 287 } 288 } 289 290 return ret; 291 } 292 293 static int submit_bo(struct msm_gem_submit *submit, uint32_t idx, 294 struct msm_gem_object **obj, uint64_t *iova, bool *valid) 295 { 296 if (idx >= submit->nr_bos) { 297 DRM_ERROR("invalid buffer index: %u (out of %u)\n", 298 idx, submit->nr_bos); 299 return -EINVAL; 300 } 301 302 if (obj) 303 *obj = submit->bos[idx].obj; 304 if (iova) 305 *iova = submit->bos[idx].iova; 306 if (valid) 307 *valid = !!(submit->bos[idx].flags & BO_VALID); 308 309 return 0; 310 } 311 312 /* process the reloc's and patch up the cmdstream as needed: */ 313 static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *obj, 314 uint32_t offset, uint32_t nr_relocs, uint64_t relocs) 315 { 316 uint32_t i, last_offset = 0; 317 uint32_t *ptr; 318 int ret = 0; 319 320 if (offset % 4) { 321 DRM_ERROR("non-aligned cmdstream buffer: %u\n", offset); 322 return -EINVAL; 323 } 324 325 /* For now, just map the entire thing. Eventually we probably 326 * to do it page-by-page, w/ kmap() if not vmap()d.. 327 */ 328 ptr = msm_gem_get_vaddr(&obj->base); 329 330 if (IS_ERR(ptr)) { 331 ret = PTR_ERR(ptr); 332 DBG("failed to map: %d", ret); 333 return ret; 334 } 335 336 for (i = 0; i < nr_relocs; i++) { 337 struct drm_msm_gem_submit_reloc submit_reloc; 338 void __user *userptr = 339 u64_to_user_ptr(relocs + (i * sizeof(submit_reloc))); 340 uint32_t off; 341 uint64_t iova; 342 bool valid; 343 344 if (copy_from_user(&submit_reloc, userptr, sizeof(submit_reloc))) { 345 ret = -EFAULT; 346 goto out; 347 } 348 349 if (submit_reloc.submit_offset % 4) { 350 DRM_ERROR("non-aligned reloc offset: %u\n", 351 submit_reloc.submit_offset); 352 ret = -EINVAL; 353 goto out; 354 } 355 356 /* offset in dwords: */ 357 off = submit_reloc.submit_offset / 4; 358 359 if ((off >= (obj->base.size / 4)) || 360 (off < last_offset)) { 361 DRM_ERROR("invalid offset %u at reloc %u\n", off, i); 362 ret = -EINVAL; 363 goto out; 364 } 365 366 ret = submit_bo(submit, submit_reloc.reloc_idx, NULL, &iova, &valid); 367 if (ret) 368 goto out; 369 370 if (valid) 371 continue; 372 373 iova += submit_reloc.reloc_offset; 374 375 if (submit_reloc.shift < 0) 376 iova >>= -submit_reloc.shift; 377 else 378 iova <<= submit_reloc.shift; 379 380 ptr[off] = iova | submit_reloc.or; 381 382 last_offset = off; 383 } 384 385 out: 386 msm_gem_put_vaddr(&obj->base); 387 388 return ret; 389 } 390 391 static void submit_cleanup(struct msm_gem_submit *submit) 392 { 393 unsigned i; 394 395 for (i = 0; i < submit->nr_bos; i++) { 396 struct msm_gem_object *msm_obj = submit->bos[i].obj; 397 submit_unlock_unpin_bo(submit, i, false); 398 list_del_init(&msm_obj->submit_entry); 399 drm_gem_object_unreference(&msm_obj->base); 400 } 401 402 ww_acquire_fini(&submit->ticket); 403 } 404 405 int msm_ioctl_gem_submit(struct drm_device *dev, void *data, 406 struct drm_file *file) 407 { 408 struct msm_drm_private *priv = dev->dev_private; 409 struct drm_msm_gem_submit *args = data; 410 struct msm_file_private *ctx = file->driver_priv; 411 struct msm_gem_submit *submit; 412 struct msm_gpu *gpu = priv->gpu; 413 struct dma_fence *in_fence = NULL; 414 struct sync_file *sync_file = NULL; 415 struct msm_gpu_submitqueue *queue; 416 struct msm_ringbuffer *ring; 417 int out_fence_fd = -1; 418 unsigned i; 419 int ret; 420 421 if (!gpu) 422 return -ENXIO; 423 424 /* for now, we just have 3d pipe.. eventually this would need to 425 * be more clever to dispatch to appropriate gpu module: 426 */ 427 if (MSM_PIPE_ID(args->flags) != MSM_PIPE_3D0) 428 return -EINVAL; 429 430 if (MSM_PIPE_FLAGS(args->flags) & ~MSM_SUBMIT_FLAGS) 431 return -EINVAL; 432 433 queue = msm_submitqueue_get(ctx, args->queueid); 434 if (!queue) 435 return -ENOENT; 436 437 ring = gpu->rb[queue->prio]; 438 439 if (args->flags & MSM_SUBMIT_FENCE_FD_IN) { 440 in_fence = sync_file_get_fence(args->fence_fd); 441 442 if (!in_fence) 443 return -EINVAL; 444 445 /* 446 * Wait if the fence is from a foreign context, or if the fence 447 * array contains any fence from a foreign context. 448 */ 449 if (!dma_fence_match_context(in_fence, ring->fctx->context)) { 450 ret = dma_fence_wait(in_fence, true); 451 if (ret) 452 return ret; 453 } 454 } 455 456 ret = mutex_lock_interruptible(&dev->struct_mutex); 457 if (ret) 458 return ret; 459 460 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) { 461 out_fence_fd = get_unused_fd_flags(O_CLOEXEC); 462 if (out_fence_fd < 0) { 463 ret = out_fence_fd; 464 goto out_unlock; 465 } 466 } 467 468 submit = submit_create(dev, gpu, queue, args->nr_bos, args->nr_cmds); 469 if (!submit) { 470 ret = -ENOMEM; 471 goto out_unlock; 472 } 473 474 ret = submit_lookup_objects(submit, args, file); 475 if (ret) 476 goto out; 477 478 ret = submit_lock_objects(submit); 479 if (ret) 480 goto out; 481 482 ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT)); 483 if (ret) 484 goto out; 485 486 ret = submit_pin_objects(submit); 487 if (ret) 488 goto out; 489 490 for (i = 0; i < args->nr_cmds; i++) { 491 struct drm_msm_gem_submit_cmd submit_cmd; 492 void __user *userptr = 493 u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd))); 494 struct msm_gem_object *msm_obj; 495 uint64_t iova; 496 497 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd)); 498 if (ret) { 499 ret = -EFAULT; 500 goto out; 501 } 502 503 /* validate input from userspace: */ 504 switch (submit_cmd.type) { 505 case MSM_SUBMIT_CMD_BUF: 506 case MSM_SUBMIT_CMD_IB_TARGET_BUF: 507 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF: 508 break; 509 default: 510 DRM_ERROR("invalid type: %08x\n", submit_cmd.type); 511 ret = -EINVAL; 512 goto out; 513 } 514 515 ret = submit_bo(submit, submit_cmd.submit_idx, 516 &msm_obj, &iova, NULL); 517 if (ret) 518 goto out; 519 520 if (submit_cmd.size % 4) { 521 DRM_ERROR("non-aligned cmdstream buffer size: %u\n", 522 submit_cmd.size); 523 ret = -EINVAL; 524 goto out; 525 } 526 527 if (!submit_cmd.size || 528 ((submit_cmd.size + submit_cmd.submit_offset) > 529 msm_obj->base.size)) { 530 DRM_ERROR("invalid cmdstream size: %u\n", submit_cmd.size); 531 ret = -EINVAL; 532 goto out; 533 } 534 535 submit->cmd[i].type = submit_cmd.type; 536 submit->cmd[i].size = submit_cmd.size / 4; 537 submit->cmd[i].iova = iova + submit_cmd.submit_offset; 538 submit->cmd[i].idx = submit_cmd.submit_idx; 539 540 if (submit->valid) 541 continue; 542 543 ret = submit_reloc(submit, msm_obj, submit_cmd.submit_offset, 544 submit_cmd.nr_relocs, submit_cmd.relocs); 545 if (ret) 546 goto out; 547 } 548 549 submit->nr_cmds = i; 550 551 submit->fence = msm_fence_alloc(ring->fctx); 552 if (IS_ERR(submit->fence)) { 553 ret = PTR_ERR(submit->fence); 554 submit->fence = NULL; 555 goto out; 556 } 557 558 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) { 559 sync_file = sync_file_create(submit->fence); 560 if (!sync_file) { 561 ret = -ENOMEM; 562 goto out; 563 } 564 } 565 566 msm_gpu_submit(gpu, submit, ctx); 567 568 args->fence = submit->fence->seqno; 569 570 if (args->flags & MSM_SUBMIT_FENCE_FD_OUT) { 571 fd_install(out_fence_fd, sync_file->file); 572 args->fence_fd = out_fence_fd; 573 } 574 575 out: 576 if (in_fence) 577 dma_fence_put(in_fence); 578 submit_cleanup(submit); 579 if (ret) 580 msm_gem_submit_free(submit); 581 out_unlock: 582 if (ret && (out_fence_fd >= 0)) 583 put_unused_fd(out_fence_fd); 584 mutex_unlock(&dev->struct_mutex); 585 return ret; 586 } 587